Controlling groups of autonomous systems

Transcription

1 Controlling groups of autonomous systems Claire Tomlin and Shankar Sastry Department of Electrical Engineering and Computer Sciences UC Berkeley

2 Planning operations for teams of aircraft Local objectives: Safety and efficiency with respect to vehicle s dynamics and actuation Global objectives: Region coverage, collision avoidance Constraints: Communication private information (Source: Prof. Robin Murphy, Univ. South Florida, Katrina Search and Rescue) Computation performed on each vehicle

3 Collision Avoidance

4 Collision Avoidance

5 Region Surveillance (Source: Prof. Robin Murphy, Univ. South Florida, Katrina Search and Rescue) Problems: Information gathering, fast decomposition of team commands into actions for each vehicle

6 Control Objectives: Automatic information gathering Safe interaction Constraints: Power budget Communication bandwidth Computational resources Secure interaction Mobile Sensor Network

7 Outline Multi-agent control Decentralized Bargaining Embedded Humans Cooperative control Safety aspects

8 Quadrotor testbed: control and software architecture Autonomous UAVs Onboard computation & sensors State and environment estimation Attitude, altitude, position and trajectory control 4 flightworthy vehicles More are being made Testbed goals Quadrotor UAV design Cooperative multi-agent control Mobile sensor networks

9

10

11 Quadrotor System Wifi Ground GPS Ground Station Computer Self Sufficient UAVs Onboard computation Onboard sensing Real Time Execution Estimation Control

12 Vehicle Design Fiberglass Honeycomb Carbon Fiber Tubing High Level Control Gumstix PXA270, or ADL PC104 Low Level Control Robostix Atmega128 Electronics Interface Tube Straps GPS Novatel Superstar II Ultrasonic Ranger Senscomp Mini-AE Inertial Meas. Unit Microstrain 3DM-GX1 Landing Gear Battery Lithium Polymer Sensorless Brushless DC Motors Axi 2208/26 Elect. Speed Cont. Castle Creations Phoenix-25

13 System Cost Metric Efficient solution Standard of comparison Marginal cost Nash Bargaining solution (NBS) NBS Standard of comparison Percentage change in marginal cost Eff.

14 Decentralized Optimization Program Vehicle 1 Penalty Weight Update Local Optimization #1 Broadcast to Neighborhood Receive Solutions Local Optimization #n Select Preferred Solution [Waslander, Inalhan, Tomlin, 2003]

15 Decentralized Optimization Program Vehicle 1 Penalty Weight Update Local Optimization #1 Broadcast to Neighborhood Receive Solutions Local Optimization #n Select Preferred Solution Proposition [Convergence to Nash Bargaining solution]: The penalty method formulation of the decentralized optimization problem converges to a solution that satisfies the necessary conditions for optimality of the Nash Bargaining solution to the centralized optimization problem. [Waslander, Inalhan, Tomlin, 2003]

16 Decentralized Optimization Program Ihalhan, Stipanovic, Tomlin, CDC, 2002

17 Four Vehicle Scenario

18 Four Vehicle Scenario

19 In flight

20 In flight

21 Embedded Humans Combine Bayesian models of human interaction with automation, and hybrid system models of automated system. Continuously evolving dynamics: spatial, temporal properties Mode switches: coupled, independent, search, cue, track New methods to learn such models from data (games, training) relay cue Displayed Information Human Decision track search Human Delay Action,t cue search Bayesian model of human s decision process track Hybrid model of automation

22 Cooperative Collision Avoidance System Requirements Scalability Safety Method for Acceleration Constrained Vehicles Each agent broadcasts its state Agents compute pairwise keep-out regions Collision avoidance control inputs mandated when violations occur

23 Cooperative Collision Avoidance System Requirements Scalability Safety Method for Acceleration Constrained Vehicles Each agent broadcasts its state Agents compute pairwise keep-out regions Collision avoidance control inputs mandated when violations occur

24 STARMAC history

25 STARMAC history

26 Pursuit Evasion Games with 4UGVs and 1 UAV ARO: Integrated Approach to Intelligent Systems (Spring 01)

27 Pursuit Evasion Games with 4UGVs and 1 UAV ARO: Integrated Approach to Intelligent Systems (Spring 01)

28 Pursuit-Evasion Game Experiment Details PEG with four UGVs Global-Max pursuit policy Simulated camera view (radius 7.5m with 50degree conic view) Pursuer=0.3m/s Evader=0.5m/s MAX

29 Pursuit-Evasion Game Experiment Details PEG with four UGVs Global-Max pursuit policy Simulated camera view (radius 7.5m with 50degree conic view) Pursuer=0.3m/s Evader=0.5m/s MAX

30 Pursuit-Evasion Game Experiment Details PEG with four UGVs Global-Max pursuit policy Simulated camera view (radius 7.5m with 50degree conic view) Pursuer=0.3m/s Evader=0.5m/s MAX

31 DARPA SEC FINAL DEMO: 2003 UCB Pursuit Evasion Games Max 20 min. games: Evader goal: get to final waypoint or avoid evader Pursuer goal: target evader Pursuer and evader restricted to same performance limits reliant on F-15 pilot s cooperation Planes on the same logical plane, but separated by 6000ft altitude at all times After first PEG, F-15 pilot request this be reduced to a 2000ft altitude separation Two scenarios: UAV as evader UAV can become pursuer

32 UCB PEG Test Plan: UAV as Evader and Pursuer UAV attempts to cross Scenario Area (SA) from East to West without being targeted by the F15, however, UAV will attempt to target F15 if suitable conditions arise UAV wins by: Reaching the END ZONE Not being targeted for 20 minutes Targeting the F15 F15 wins by targeting the UAV Note: F15 performance is restricted

33 Flight Test: Results

34 Flight Test: Results

35 NEST Demo Movie

36 NEST Demo Movie

37 Hunt Research Plans What is missing to date Multi-person, multi-objective games Biological inspiration drawn from hunting patterns Group/Team behavior modeling Research Strategy on HUNT Solution Concepts for Partially Observable Games with Partial Information Existence of Solutions: heuristics drawn from predator prey interactions in the wild: team behaviors New S&T to be developed on HUNT Traditional solutions to games are causal : that is to say plans depend on observations: this excludes the kind of predictive team behavior that characterizes hunting. New predictive solution concepts are needed. Learning of Utility Functions of Predator-Prey from engagement to engagement. Traditional learning has focused on updates of strategy rather than learning of utility functions of the adversary